Research On Neural Network Based Fault Location Methods For Satellite Attitude Control System

Attitude control system(ACS)is one of the most critical subsystems of the spacecraft and also one of the subsystems with high incidence of failure.It is of great significance to study the fault diagnosis(FD)technology of ACS to ensure its operation is stable and reliable.ACS is a typical nonlinear system with many components,complex structure and functions,and its operation environment if full of unknown factors.It is a challenge to design a reasonable and effective FD method.The fault location(FL)of actuator and sensor in ACS control loop is studied.In a case of sensor and actuator fault of a high-precision spacecraft,of which the dynamics model is complex,the component fault is propagating in a closed loop.Based on the available state measurement information and instruction information of the ACS,an FL scheme based on neural networks(NN)and support vector machine(SVM)is proposed for actuator and sensor fault.Firstly,based on the ACS mechanism,the FL logic of actuator and sensor fault is constructed.Then,an NN is applied to model the ACS dynamics.By constructing the dynamics observer and kinematics observer,the dynamics process and kinematics process are observed respectively,residuals are generated,and residual features are extracted,and the fault detection is realized by SVM.Finally,based on the fault detection results and the FL logic,the FL of actuator and sensor in closed-loop ACS is realized without using model parameters.A simulation is carried out on a semi-physical simulation platform,and the results show that the proposed method can effectively locate actuator and sensor fault.Transfer learning(TL)is introduced to improve the FL method for actuator and sensor.Aiming at the problem that the ACS has few fault samples,based on TL,The NN and SVM is trained with data generated in a nominal model simulation,which constructed the dynamics observer and kinematics observer for the nominal model.Then the available healthy spacecraft data is used to finetune the pretrained NN to obtain the dynamics observer for the spacecraft.Compared with the original method,the improved method does not require fault samples of the spacecraft,only a small number of healthy samples is required,and the computation cost is reduced,which makes the method more practical.The FL of spacecraft with multiple attitude sensors concurrent faults is studied.Based on the attitude determine(AD)principle,the analytical model of attitude sensors,and the available healthy ACS sensor output,multiple observers are constructed,taking advantage of the compression and decompression functions of the autoencoder(AE)NN.The fault signals can not be fully reconstructed during compression and decompression,so the residuals are generated.To improve the accuracy and reliability of fault detection,residual features are extracted and input into a softmax classifier for training,which established a fault classifier.Finally,the fault detection results of the multiple observers are voted to realize the multiple sensors fault location.The proposed method does not require the sensor deployment parameters,and it is not affected by the sensor deployment errors.The simulation results show that the proposed method can effectively detect and locate single fault and multiple fault.